Persistent Poverty and Children's Cognitive Development:
evidence from the UK Millenium Cohort Study
Andy Dickerson
Gurleen Popli
ISSN 1749-8368
SERPS no. 2011023
Originally Published: November 2011
Updated: December 2014
Persistent Poverty and Children’s Cognitive
Development: Evidence from the UK Millennium
Cohort Study
Andrew Dickerson and Gurleen K. Popli
Department of Economics
University of She¢ eld
December 2014
ABSTRACT
We use data from the four sweeps of the UK Millennium Cohort Study
(MCS) of children born at the turn of the century to document the impact
that poverty, and in particular persistent poverty, has on their cognitive development in their early years. Using Structural Equation Modelling (SEM),
we show that children born into poverty have signi…cantly lower test scores
at age 3, age 5 and age 7, and that continually living in poverty in their early
years has a cumulative negative impact on their cognitive development. For
children who are persistently in poverty throughout their early years, their
cognitive development test scores at age 7 are almost 20 percentile ranks lower
than children who have never experienced poverty, even after controlling for
a wide range of background characteristics and parental investment.
Keywords: child poverty, cognitive development
JEL classi…cation codes: I32, J13, J62
Address for correspondence: Gurleen Popli, Department of Economics, University of
She¢ eld, 9 Mappin Street, She¢ eld, S1 4DT, UK. Email: g.popli@she¢ eld.ac.uk. T:
+44 (0)114 222 3485. F: +44 (0)114 222 3458.
Persistent Poverty and Children’s Cognitive Development:
Evidence from the UK Millennium Cohort Study
"Give me a child until he is seven and I will give you the
man." (attrib.) St. Francis Xavier (1506 –1552)
1
Introduction
In this paper we investigate the impact of persistent poverty on the cognitive
development of children in the very early years of their lives. We use the
UK Millennium Cohort Study (MCS) which is a sample of 19,000 children
born in the UK around the turn of the century. We trace their cognitive
development as measured in a series of standard tests up until they are 7
years old. Our focus is on the impact of living in poverty on their cognitive
development. We assess the impact of both episodic (period-by-period)
poverty and persistent poverty, in order to examine the cumulative impact
of multiple and continuous periods of deprivation.
It has become increasingly apparent that there is a strong link between
children’s development and educational attainment, and their family background (Blanden et al., 2007; Feinstein, 2003; Heckman and Masterov, 2007;
Gregg and Macmillan, 2009). Speci…cally, there is a large literature exploring the impact of poverty and low income on the development of children.
Brooks-Gunn and Duncan (1997) review evidence from numerous national
longitudinal data sets for the US (such as the Panel Study of Income Dynamics (PSID), National Longitudinal Study of Youth (NLSY), and Children of
NLSY) focusing on the consequences of poverty across a range of outcomes
for children, and the pathways through which poverty might operate. Much
of the evidence they describe points towards the negative impact of poverty
on child development.
The link between early educational attainment and socio-economic status
(SES) of families has also been emphasised by policy makers. Coupled
with his administration’s emphasis on education, in 1999 Prime Minister
Tony Blair also famously pledged to end child poverty ‘within a generation’:
“Our historic aim will be for ours to be the …rst generation to end child
poverty,” (Tony Blair, Beveridge Lecture, 1999). This commitment was
accompanied by a range of reforms and initiatives designed to tackle what
had become a crisis, with one quarter of all children in Britain living in
poverty in 1999. The following decade saw signi…cant reductions in the child
poverty rate measured in both relative and absolute income terms (Waldfogel,
2010). However, the rate of progress slowed somewhat after 2003/04, and
1
progress towards the intermediate target of cutting the child poverty rate
in half by 2010 was missed. Subsequently, the prospect of ‘eliminating’
child poverty by 2020 looks increasingly di¢ cult. However, in March 2010,
the UK Child Poverty Act enshrined in law the commitment to end child
poverty by 2020. Explicit targets have been set in terms of relative income,
material deprivation and absolute income measures. In addition, and as
a late addition to the legislation, the signi…cance of ‘persistent poverty’was
recognised, with a target to prescribed by regulation before 2015 (at the time
of writing, this target was yet to be set).
More recently, the Field review (Field, 2010) on ‘Poverty and Life Chances’
called into question the focus on income poverty. Instead, the review recommended greater attention be paid to the problem of the intergenerational
transfer of poverty. The role of family background, the quality of parenting,
and children’s opportunities for learning and development were argued to
be crucially signi…cant in determining adult outcomes because of their importance in children’s development before age 5 (‘Foundation Years’in the
terminology of the review). Consequently, the review argued that supporting
parents and their children in these early years through directed government
spending should be the priority. Thus, rather than making income transfers
to poorer families through the tax credit system, the Field review recommends that consideration should always be given to whether that income
would be more e¤ectively spent in improving early years provision of services
such as Sure Start (a programme similar to Head Start in the US) in order
to improve the outcomes that children from poor families might achieve in
their adult lives. Such recommendations represent a clear change in focus.
The impact of persistence poverty remains a largely unexplored aspect of
the importance of family background and other characteristics on children’s
cognitive development and educational attainment. Our paper is a contribution towards an investigation of this important issue. Speci…cally, we examine the relative importance of both family background (including parental
investment) and income poverty – especially the persistence of poverty –
for children’s early cognitive development. This is one of the few papers
to systematically and robustly examine the impact of persistent poverty on
young children’s cognitive development in contemporary Britain. Using a
Structural Equation Modelling (SEM) approach, we show that children living
in poverty have signi…cantly lower cognitive test scores, even after controlling for a wide range of background characteristics and parental investment;
and that the legacy of persistent poverty in their early years is a cumulative
negative impact on their cognitive development.
The remainder of this paper is organised as follows. In the next section,
we brie‡y review the relevant literature on cognitive development, family
2
background and poverty. Section 3 presents the method that we employ to
estimate the link between child poverty and cognitive development. Section 4
describes the data, and the tests that are used to measure children’s cognitive
development. Section 5 presents our main results, and section 6 draws some
conclusions and implications.
2
Background literature
When discussing the impact of poverty on children, the timing and the duration of poverty is often highlighted (Duncan et al., 1998). There is growing
evidence across various disciplines (neuroscience, development psychology,
and economics among them) that the environment in the early years of a
child’s life has a signi…cant impact on their development and ability formation (Knudsen et al., 2006). Thus it is not surprising that the impact of
poverty is also found to be greater for poverty experienced in early childhood
(usually de…ned as birth to 7 years) relative to late childhood or adolescence.
Duncan et al. (2010) use the PSID and …nd signi…cant and ‘quantitatively
large detrimental e¤ects’of poverty in the early years (birth to 5 years) on a
range of adult outcomes (earnings and hours worked). The theoretical explanation for these larger e¤ects from early exposure of poverty comes from the
‘self-productivity’argument given by Heckman and Masterov (2007), where
development in later years is dependent upon development in early years.
Evidence also suggests that long exposure to low resources is more detrimental than transitory changes in family fortunes. Carneiro and Heckman
(2002) use US data to study the relationship between family income and
schooling of children. Their …ndings suggest that it is long-term factors such
as better family resources throughout the child’s formative years, rather than
short-term liquidity constraints, that largely account for the family income
gap in college enrolment. Findings by Cameron and Heckman (2001), again
based on US data, also demonstrate that it is long-term in‡uences associated with parental background and parental income throughout the child’s
adolescent years that largely account for the racial-ethnic college enrolment
di¤erential. Using data from Indonesia, Pakpahan et al. (2009) …nd that
children who grow up in chronically (persistently) poor households have a 31
percentage point higher risk of continuing to live in poverty as adults relative
to children from non-chronically poor households.
One key challenge in identifying the link between poverty and child development is in disentangling the e¤ect of poverty from a range of factors
associated with poverty which in themselves have a negative impact on children’s development. For example, children in poor households often also
3
have young, less educated, and single mothers. Each of these factors (young
mother, less educated mother, single mother) by itself is associated with
poorer outcomes for children (Mayer, 1997). Plug and Vivjerberg (2005)
use adoptee data and show family income has a signi…cant impact on schooling outcomes; adoptees with access to better family income have better educational outcomes. Dahl and Lochner (2012) use an instrumental variable
approach to establish causality from family income to child development;
even after controlling for numerous confounding factors they …nd that family
income has an independent causal link to child development, more so for the
low-income families.
There are a number of pathways via which poverty can impact child
development: health and nutrition; home environment; parental interactions
with children; parental health; neighbourhoods etc. (Corcoran, 1995; Duncan
et al., 1998). An important pathway often stressed in the literature is the
home environment, which is taken to include everything from the quality and
quantity of time inputs by the parents, to the quality and quantity of goods
(learning resources, toys, etc.) inputs provided by the parents to the child
(Leibowitz, 1974). The home environment is often viewed as a mediating
factor for most of the external factors impacting the children, such as the
government programs or low income (Becker and Tomes, 1986; Todd and
Wolpin, 2003; Gelber and Isen, 2013). Using the data for the US from the
PSID and the Infant Health and Development Program, Brooks-Gunn et al.
(1993) show that provision of learning experiences in the home can account
for up to half of the e¤ect of poverty on the IQ scores of …ve year olds.
Similarly, a more recent study by Gelber and Isen (2013) analysing the Head
Start program from the US …nds that a signi…cant part of the impact of the
program on child development was via the increased parental investment in
the child, as a result of participation in the program.
In our analysis we focus on the early years of children’s development,
as re‡ected in their cognitive development. We estimate the impact of
both episodic (short term) and persistent (long term or chronic) poverty.
We examine both the direct and the indirect impact of poverty on child
development; for the indirect impact we speci…cally explore the pathway of
home environment (which we call ‘parental investment’). Finally, we also
address two key empirical issues of measurement error and endogeneity of
parental inputs.
First is the issue of measurement error, both in estimating cognitive ability and in measuring the parental investment in the child. In our framework,
we assume that both the true ability of the child and the true investment in
the child cannot be observed. Instead, what we have are a range of (imperfect) measures of cognitive ability and parental investment. Consider …rst
4
the unobserved (latent) cognitive ability of the child. What we observe are
test scores which are correlated with latent cognitive ability, but measure it
with error. Cunha and Heckman (2008) discuss the issue in terms of ‘measurement error’while Jerrim and Vignoles (2013) focus on the implications
in terms of ‘regression to the mean’. The basic argument is that the imperfection/randomness in testing means that classifying children as high or low
ability on the basis of a single test is liable to be subject to error since getting
a relatively high (low) score on a given day is likely to be followed by a less
extreme score (i.e. will be lower (higher)) if they were tested on another
day. To mitigate this problem we can use multiple tests at each age to estimate the latent cognitive ability of the child. A similar issue arises with the
measurement of parental investment. However, we have numerous proxies
available in our data which are related to the latent parental investment in
the child.
Second is the issue of endogeneity of inputs, especially parental investment
(see Todd and Wolpin, 2003, 2007, for details). The source of the problem
is that there are inputs we do not observe but parents do, and that parents
may modify their inputs based on what they observe of the child, leading to
reverse causation. In our estimation framework we explicitly account for this
potential endogeneity of inputs.
Next we brie‡y review studies using the UK data and put our work in
the context of the UK speci…c literature.
2.1
Review of UK studies
For the UK, using data drawn from the 1970 British Cohort Study, Feinstein
(2003) showed that parental SES has an important and long-lasting impact
on children’s development and attainment. While early cognitive development is a good predictor of educational quali…cation attainment 20 years
later, children from low SES families are particularly disadvantaged. He
also argued that children in low SES families are less likely to demonstrate
high early scores, and even if they do show signs of good initial cognitive
development, this advantage is soon eroded. Any upward mobility of children with low initial attainment is for children from medium and high SES
families. Our paper has a number of parallels with Feinstein’s study in that
we are also interested in the impact of family background on children’s early
cognitive development, although our focus is on poverty and the persistence
of poverty rather than di¤erences over time by social status.
Gregg and Macmillan (2009) examine the impact of parental income on
children’s education and test scores (the youngest children they have are aged
7 years) using various UK cohort studies: National Child Development Study
5
(born in 1958); British Cohort Study (born in 1970); three separate cohorts
constructed from the British Household Panel Survey (for children born in the
late 1970s, early 1980s, and late 1980s); Avon Longitudinal Study of Parents
and Children (a Bristol-based birth cohort of children born in 1991/92); and
the Longitudinal Study of Young People in England (a national sample of
children born in England in 1989/90). They consistently …nd that children
born into poorer families have a lifelong disadvantage.
Goodman and Gregg (2010) utilise the second and the third sweeps of
the MCS. Their focus is on explaining the rich-poor gap in the cognitive
ability of children by analysing the in‡uence of aspirations and behaviour of
parents on the outcomes of their children. However, they do not take into
account the persistence of poverty in documenting or explaining the existence
of the gap in ability. Blanden and Machin (2010) also utilise the second
and the third sweeps of the MCS. They examine the connection between
parental income and children’s vocabulary and behaviour. Consistent with
the previous literature, their …ndings also suggest that better child outcomes
(in terms of vocabulary development and behaviours) are associated with
higher income. For example, children from families in the top quintile in
terms of income are more than one year ahead in vocabulary development at
age 5 as compared to children from the bottom quintile.
None of the studies cited above examine the impact of persistent poverty.
In contrast, Schoon et al. (2010) use the MCS data to look at the impact
of persistent …nancial hardship (measured as the family being in receipt of
state bene…ts) on the cognitive and behavioural development of children at
age 5. Their …ndings suggest that persistent …nancial hardship has a large
and negative impact on children’s cognitive development, while the impact
on children’s behavioural adjustment is rather less. Further, this negative
impact is mitigated by the ‘protective factors’in the family environment. In
a related paper using the same data, Schoon et al. (2012) examine the impact
of persistent (income) poverty and ‘family instability’(de…ned as changes in
mothers’relationship status: married, cohabitating, or single) on children’s
cognitive ability. The results from this paper con…rm their earlier …ndings
and further illustrate that, after controlling for poverty, family instability has
no signi…cant association with the cognitive development of children.
Kiernan and Mensah (2009) also use the MCS data and investigate the impact of persistent poverty, maternal depression, and ‘family status’(de…ned
as mothers’ relationship status) on the cognitive and behavioural development of children, at age 3. Their …ndings also suggest that poverty has a
negative impact on the development of children and, once poverty is taken
into account, the e¤ects of both maternal depression and family status are
weak. In a related paper, Kiernan and Mensah (2011) look at the impact of
6
parenting and persistent poverty on cognitive development of children at age
5. Their …ndings echo those of Schoon et al. (2012): the negative impact of
persistent poverty is mitigated by positive parenting.
Our paper is di¤erent from Schoon et al. (2010, 2012) and Kiernan and
Mensah (2009, 2011) in number of important aspects. First, we explicitly
address the issues of measurement error and endogeneity of inputs. Second,
we consider a longer time horizon by examining children’s development at
age 3, age 5 and age 7. This gives us an important advantage in modelling
the persistence in cognitive development, and also allows us to include a
period when the children have been attending school. Third, we explicitly
model the ‘parental investment’in children; the so called ‘protective factors’
in Schoon et al. (2010), and the ‘index of parenting’in Kiernan and Mensah
(2011).
Barnes et al. (2010) examine the impact of persistent poverty (de…ned as
being in poverty for at least 3 of the 4 annual interviews of the Growing up
in Scotland study) on young children in Scotland. They note that poverty is
multi-dimensioned and that many, if not most, of its e¤ects can be captured
through correlated characteristics such as low parental education, poor health
etc. Indeed, low income is not statistically signi…cantly correlated with child
outcomes (such as being overweight, poor language, social and emotional
development etc) once all of these other family and various environmental
factors are taken into account. Of course, this does not mean that income
is not important for child outcomes, but rather that its impact is indirect,
through its e¤ect on other factors which are correlated with outcomes. In
our analysis, we are able to capture and distinguish between both the direct
and the indirect impact of income poverty.
3
Estimation method
In our framework, we adopt a value added plus lagged inputs model of ability
formation (Todd and Wolpin, 2007), where child’s current cognitive ability
depends on the previous ability of the child and the past inputs (parental
investments). Further, as in Cunha and Heckman (2008) we assume both the
child’s cognitive ability and the parental investment in the child are latent.
The true ability of the child and the true investment in the child cannot
be observed. Instead what we have are a range of (imperfect) measures
of cognitive ability and parental investment. So, for example, the reading
test score is just one measure of a child’s cognitive ability; similarly a parent
reading to the child is just one measure of the parental investment in the child.
To understand the link between poverty and the cognitive development of
7
the child, mediated by parental investment, we use SEM. The SEM has two
components –a structural model and a measurement model.
3.1
Structural model
Let t be the stock of latent cognitive skill (ability) of the child at time t.
A child’s ability at time t, t , depends on past ability stock, t 1 , and past
parental investment, t 1 ; it also depends on some exogenous covariates Xt ,
poverty being one such covariate. Evolution of ability over time is thus given
by:
t
=
1t t 1
+
2t t 1
+
3t Xt
+
t
(1)
where t = 1; : : : T , represent the di¤erent time periods of childhood, with
t = 0 representing the initial endowments that a child is born with; 1t , 2t ,
and 3t are time-varying parameters to be estimated; and t is the normal
error term, assumed to be independent across individuals and over time.
Parental investment is also assumed to be latent, and is in‡uenced by
some (exogenous) covariates, Xt (including poverty),
=
t
4t Xt
+
(2)
t
where 4t is the vector of time-varying parameters to be estimated; and t
is the normal error term, assumed to be independent across individuals and
over time.
For period t = 1 equation (1) will be: 1 = 11 0 + 21 0 + 31 X1 + 1 . In
our empirical exercise we do not have speci…c measures to separately identify
the initial endowments, 0 , and initial parental investment in the child, 0 ,
and hence we assume that these together depend in a linear fashion on a set
of covariates, X0 . So for t = 1 we estimate:
1
3.2
=
11 X0
+
31 X1
+
1
(3)
Measurement model
As both ability and parental investment are taken to be latent, we have a
measurement model for each of them.
Yj;t =
j;t
+
j;t t
+ "j;t
(4)
Yj;t =
j;t
+
j;t t
+ "j;t
(5)
8
k
where Yj;t
(for k 2 f ; g and j 2 f1; :::; mkt g) are the measures available
for the latent ability and latent parental investment at time t. mkt are the
number of measures available, such that mkt
2. kj;t are the factor loadings
which can be interpreted as the amount of information that the measures
k
contain about the latent variables ( t and t ). "kj;t are the measurement
Yj;t
errors, which capture the di¤erence between the observed measures and the
unobserved latent variables. kj;t and kj;t can depend on regressors as long
as they are independent of the latent variables ( t and t ) and the error term
("kj;t ).
3.3
Identi…cation
The factor loadings, in equations (4) and (5), can be identi…ed only up to
a scale, so we need to normalize them; the normalization we use here is:
Further, we can not separately identify the mean of the
1;t = 1.
1;t =
latent variable, E( ), and the intercepts j;t ; we need to normalize one of
them, we assume E( ) = 0 and identify j;t . Similarly we assume E( ) = 0
and identify j;t .
To be able to identify all the parameters of interest in equations (1) to
(5) we need following assumptions:
Assumption 1 : "kj;t is mean zero and independent across agents and over time
for t 2 f1; : : : T g; j 2 f1; :::; mkt g; and k 2 f ; g.
Assumption 2 : "kj;t is mean zero and independent of ( t ; t ) for t 2 f1; : : : T g;
j 2 f1; :::; mkt g; and k 2 f ; g.
Assumption 3 : "kj;t is mean zero and independent from "li;t for t 2 f1; : : : T g;
j 2 f1; :::; mkt g; and l; k 2 f ; g such that l 6= k.
In the empirical analysis to aid computation we further assume that "kj;t ,
and t have a normal distribution, though this is not needed for identi…cation.
The assumptions we make here are the same assumptions as made by Cunha
and Heckman (2008, page 747 and Appendix A2); see their paper for further
details.
3.3.1
Identi…cation of the factor loadings
Consider t . Assume for simplicity that mt = 2, i.e. we have only two
measures for t : f[Yj;t ]2j=1 gTt=1 . From the data available we can calculate the
covariance between the di¤erent measures, which gives us the following set
of equations (recall the normalization 1;t = 1):
Cov(Y1;t 1 ; Y1;t ) = Cov(
9
t 1; t)
(6)
Cov(Y2;t 1 ; Y1;t ) =
Cov(Y1;t 1 ; Y2;t ) =
2;t 1 Cov( t 1 ; t )
(7)
2;t Cov( t 1 ; t )
(8)
can be identi…ed by taking the ratio of equation (7) to equation (6),
and 2;t can be identi…ed by taking ratio of equation (8) to equation (6).
Similarly we can identify the factor loadings j;t for the latent parental
investment, up to the normalization 1;t = 1, by exploiting the covariances
between the measures of parental investment Yj;t .
2;t 1
3.3.2
Identi…cation of the structural parameters
Once the factor loadings have been estimated we can use two-stage least
squares to estimate the structural parameters. In the …rst-stage we take the
weighted average of the measures to get the error-corrected estimates of the
latent variables. For example, consider t , we can use the estimated factor
loadings to construct:
bt =
mt
X
! j;t Yj;t
where
j=1
b j;t
! j;t = P
mt
j=1
2
b j;t
2
(9)
where bt is the error-corrected estimate of true latent ability t . We can similarly construct bt , the error-corrected estimate of true parental investment
In the second-stage bt , bt 1 and bt 1 can be substituted in equations
t.
(1) to (3) to estimate the structural parameters. For details refer to the
discussion in Cunha (2011).
3.4
Endogeneity
If we assume that t is independent of t , then assumptions in section 3.3
fully identify the model. We call this as the ‘baseline model’ in our empirical analysis. However, endogeneity of inputs is a concern: one way to think
about this is reverse causation where parents observe some aspect of child’s
current ability which has an impact on their investment in the child, so that
t can be expected to impact t . In equation (1) using lagged parental investment ( t 1 ) to explain child’s current ability ( t ) should solve the issue
of reverse causation, as t cannot impact t 1 (referred to as the value added
with lagged inputs speci…cation by Todd and Wolpin, 2007). However, we
have a dynamic model where t 1 impacts t and if the inputs are endoge10
neous then t 1 can also impact t 1 . This will then violate the assumption
that t is independent of t .
To address the issue of endogeneity of inputs in the most general way we
specify a parental investment function, as suggested by Cunha et al. (2010):
t
=
1t t
+
2t Xt
+
3t Rt
+
t
(10)
where t is the normal error term that is orthogonal to ( t , Xt , Rt ); and we
assume that there exists at least one such variable, Rt , that impacts parental
investment but not the ability of the child. Rt can be interpreted to re‡ect
the family resources or constraints that limit the ability of the parents to
invest in their children, but do not have a direct impact on child’s ability.
Cunha et al. (2010) discuss in detail the justi…cation of the investment function given by equation (10); they further suggest that to estimate equation
(10) we can use two-stage least squares where the past values of Rt are used
as proxies for t .
3.5
Direct versus indirect e¤ects
The SEM approach allows us to capture and identify both the direct and the
indirect e¤ects of poverty on cognitive development. The direct e¤ects are
simply how poverty a¤ects cognitive development, while the indirect e¤ects
capture how poverty a¤ects parental investment and parenting style, which
in turn impact upon cognitive development. Equation (1) gives us the
direct impact of the exogenous variables (including poverty) on cognitive
skills while equation (2) gives us the direct impact of exogenous variables
on parental investment. Equations (1) and (2) together give us the indirect
e¤ects of the exogenous variables on cognitive skills through their impact on
parental investment. Separately identifying the direct and indirect e¤ects
allows us to compute the total e¤ect of each of the exogenous variables on
children’s cognitive development. We therefore identify three e¤ects: (i) the
direct impact of parenting inputs on children’s ability, (ii) the direct impact
of poverty on children’s ability, and (iii) the indirect impact of poverty on
children’s ability, via its impact on parenting inputs.
3.6
Episodic versus persistent poverty
We estimate two di¤erent speci…cations using SEM. In the …rst speci…cation,
in vector Xt in equation (1), we include a (1, 0) dummy only for the current
poverty status (Pt ). This captures the direct impact of episodic poverty on
cognitive development ( t ). Implicitly, it is assumed that the previous periods
11
of poverty do not have any direct impact on current cognitive development.
Any previous poverty episodes (Pt i ) only have an indirect impact on current
cognitive development, via lagged cognitive development ( t 1 ), and lagged
parental investment ( t 1 ). Similarly, Xt only includes Pt .
To capture the impact of persistent poverty on the child’s cognitive development, in the second speci…cation we include dummies for all past poverty
states in equation (1); i.e. Xt now includes Pt , Pt 1 , Pt 2 , etc.. This allows
us to identify the direct impact of persistent poverty on the cognitive development of the child by cumulating the estimated coe¢ cients of Pt , Pt 1 , Pt 2 ,
etc.. The motivation for this speci…cation comes from the wider literature
on poverty which makes the case that the e¤ect of a period of poverty on
an individual is likely to be di¤erent depending on whether this period of
poverty was preceded by poverty or relative a- uence (i.e. not in poverty);
see Foster (2009), Bossert et al. (2012) and Dutta et al. (2013). This speci…cation thus allows us to distinguish between the direct e¤ects of episodic
and persistent poverty, where the latter has often been called the ‘scarring’
impact of poverty in the literature. The indirect impact of previous episodes
of poverty via lagged cognitive development and lagged parental investment
still exists. Similarly we also allow parental investment to be in‡uenced by
persistent poverty; Xt now includes Pt , Pt 1 , Pt 2 , etc.
A diagrammatic representation of the estimated structural and measurement model is given in Figure 1. Estimation is performed using Mplus v7
(Muthen and Muthen, 2010).
4
Data and measurement
The UK Millennium Cohort Study (MCS) is following a large sample of
around 19,000 babies born in 2000-01. The …rst sweep (MCS1) took place
in 2001-02 when these babies were, on average, around 9 months old, and
recorded details of their family background, mothers’pregnancy and birth,
and the early months of their lives. The second sweep (MCS2) took place
when the children were around 3 years old, while the third sweep (MCS3) was
administered when the children had reached age 5 and had started school.
Finally, the fourth sweep (MCS4) was undertaken in 2008 when the children
were 7 years old. The age 11 survey, the …fth sweep, started in 2012. In
our analysis we use the …rst four sweeps of the MCS.
Information is gathered in face-to-face interviews on a wide range of socioeconomic and demographic characteristics about the child, their family (parents and grandparents), parenting activities, cognitive assessments, and early
education. The survey has a clustered strati…ed design, with oversampling
12
of: ethnic minorities (Asian and Black families); children living in disadvantaged areas; and children from the three smaller countries of the UK
(Scotland, Wales and Northern Ireland). Weights which take account of
di¤erential sampling, non-response and attrition have been used throughout
the analysis –see Hansen (2012) for details.
4.1
Measuring child poverty
Household income data in the MSC is gathered using a banded question
recording total net income from all sources. A large number of income
bands are used (19 in MCS4 for example) with the top and bottom bands
open-ended. Di¤erent sized bands were used for lone parent households, as
compared to couples, given their lower household incomes in general. Rather
than using the mid-point of the reported band as an estimate of household
income, a continuous income measure is imputed using interval regression
(Stewart, 1983), with predictors including age, labour market status, region,
bene…t recipient, ethnicity, highest education level, housing tenure and number of children. This imputation also facilitates predicting household income
for non-respondents (either ‘refusal’or ‘don’t know’, representing about 10%
of households in each sweep).
While there are several criteria currently in use for measuring child poverty,
a commonly utilised measure is relative income poverty. This is the measure of poverty that is reported in the o¢ cial Households Below Average
Income (HBAI) statistics on poverty (HBAI, 2010), and is de…ned as living
in a household with net equivalent income less than 60 percent of median
UK household income. Equivalisation takes into account family size and
composition by rescaling income by the number of ‘equivalent adults’. As
with the HBAI, the MCS uses the modi…ed OECD household equivalence
scale (OECD, 2009) which weights the …rst adult as 0.67, the second adult
and each child over 14 as 0.33, and each child under 14 as 0.20. The MCS
equivalised income is then compared to the o¢ cial poverty thresholds from
the HBAI for the appropriate year of the MCS sweep. Children living in
households below the HBAI threshold for that year are de…ned as being in
poverty for that sweep. We use this measure of child poverty, made available
by MCS, in our analysis.
4.2
Cognitive test scores
It is di¢ cult to capture the latent cognitive ability of a child at age 9 months
(i.e. in MCS1) since there are no tests for cognitive ability for children that
young. What we can measure at that age is their development –i.e. their
13
physiological and psychological functioning –and in particular, whether they
have reached particular age-speci…c ‘developmental milestones’ that most
children can do at their age. The MCS uses the well-established Denver
Developmental Screening Test (Frankenburg and Dodds, 1967; Frankenburg
et al., 1992) with assessment based on the responses given by the main respondent. The version of the test used in MCS assesses children in three
areas: …ne motor function (picking, passing, etc.); gross motor function (sitting, standing, etc.); and communicative gesture (smiles, nods for yes, etc.)
(Dezateux et al., 2004; Schoon et al., 2012). A child is classi…ed as having
a delay in a particular item if s/he is unable to perform a task that 90% of
the children of their age can do.
The MCS records a number of standard tests of cognitive development
at ages 3 (MCS2), 5 (MCS3) and 7 (MCS4) years. In each case, these are
age-appropriate tests administered to the children themselves. We focus on
the children’s performance across all of these tests since they re‡ect di¤erent
cognitive abilities and educational concepts and performance, and provide
di¤erent indicators of ability. There are two tests in MCS2, and three in each
of MCS3 and MCS4. These tests are described brie‡y below while Connelly
(2013) and Hansen (2012) provide full details of the implementation of these
tests in the MCS.
The British Ability Scales (BAS) are a set of standard age-appropriate
individually-administered tests of cognitive ability and educational achievements suitable for use with young children –see Elliott et al. (1996, 1997) for
further information. Six di¤erent BAS tests have been administered across
the MCS sweeps. The di¢ culty of the tests is changed dynamically to re‡ect
the child’s performance in the initial set of items in order to present the child
with the most appropriate test for their ability. The BAS Naming Vocabulary test (BAS-NV) assesses expressive verbal ability and vocabulary, as well
as language development. The children are shown a series of coloured pictures of objects one at a time which they are asked to name. The raw scores
are then adjusted to take account of the di¢ culty of the item set administered and the age of the child (in 3 month bands) using a set of standard
look-up tables. This test was administered in MCS2 and MCS3. In the
BAS Pattern Construction test (BAS-PC), the child replicates a design by
putting together ‡at squares or solid cubes with black and yellow patterns
on each side. The child’s score is based on both speed and accuracy in the
task. This test assesses spatial problem solving, but also dexterity and coordination. Once again, the raw scores are adjusted for age and the di¢ culty
of the test items with reference to a set of standard tables. This test was
administered in MCS3 and again in MCS4. The BAS Picture Similarity test
(BAS-PS) was administered in MCS3. This test assesses non-verbal reason14
ing or problem solving: the child is shown 4 pictures and asked to identify
a further related/congruent picture. Finally, in the BAS Word Reading test
(BAS-WR) which was administered in MCS4 (age 7), the child reads a series
of words presented on a card. This assesses the child’s educational knowledge
of reading.
In addition to the six BAS-based tests, two further tests were administered. First, in MCS2, the Bracken School Readiness Assessment (BSRA)
is used to assess the ‘readiness’ of young children for formal education by
testing their knowledge and understanding of a range of basic concepts –see
Bracken (2002). MCS2 employs six of the subtests which speci…cally evaluate: colours; letters; numbers/counting; sizes; comparisons; and shapes.
The BSRA test result is a composite score based on the total number of correct answers across all six subtests. Second, in MCS4, children’s numerical
and analytical skills are assessed using a variant of the National Foundation
for Educational Research (NFER) standard Progress in Maths (PiM) test in
which a range of tasks covering number, shape, space, measures and data are
assessed.
For each of the tests, we use the age-standardised scores, and construct
the child’s percentile ranking across all children in the MCS who complete
the test to take account of di¤erences in scale and dispersion between the
tests. The percentile rankings record on a scale of 0 to 100 the percentage of
children in the sample completing the test who are ranked below the child’s
score. Thus a child’s ranking of 90 on a particular test indicates that 90
percent of children scored lower in the test; the child is thus in the top
10% of the speci…c test score distribution. Percentile rankings also provide
a convenient and informative metric against which to record the in‡uence of
poverty on the di¤erent cognitive skills assessed in each of the tests.
4.3
Independent variables
There is considerable evidence in the literature that children’s cognitive outcomes are in‡uenced by the SES and other characteristics of their family,
including parents’ (especially mothers’) education and family structure, as
well as parental investment (see, for example, Field, 2010; Ermisch, 2008;
Kiernan and Huerta, 2008; Kiernan and Mensah, 2011; Schoon et al., 2010,
Melhuish et al., 2008). Thus, we include a range of variables in our empirical model which may impact on children’s cognitive development. Identical
questions are not asked in every sweep of the MCS since the focus is on
making the survey questionnaire age-relevant.
15
4.3.1
Parental investment
The measures used for capturing parental investment are motivated by the
Melhuish et al. (2008) paper; where the authors discuss the di¤erent home
learning variables (like reading to the child) and the variables that capture
the social/routine activities (like regular bed time) which constitute the home
environment of (input into) the child. The measures used here are similar
to those in the ‘HOME’score, often used in the US based studies to capture
inputs into the child; while some of the measures are directly linked to the
cognitive development of the child (like reading to the child), the other are
thought to provide an environment conducive to learning (like regular bed
time); see Todd and Wolpin (2007), for further discussion and references.
A range of variables in the MCS record di¤erent dimensions of the ‘home
learning environment’(HLE) and social/routine activities. For the former,
these include: how often the child is read to (5 categories from ‘never’ to
‘every day’); how often the child paints or draws at home (5 categories);
how often the child is helped with reading (5 categories); how often the
child is helped with writing (5 categories); how often the child is helped with
maths (5 categories); and how often the child visits the library (3 categories).
In addition, fathers are also asked how often they read to their child (5
categories). These represent 6 of the 7 dimensions that Melhuish et al.
(2008) utilise in constructing their HLE index for the E¤ective Pre-school,
Primary & Secondary Education (EPPSE) data although here we take a
latent variable approach rather than aggregating the scores on each dimension
into a single numerical index as in Melhuish et al. (2008).
In addition to the home learning environment, a number of variables
re‡ect parenting ‘style’, including socialisation, routine and discipline. In
MCS1, when the children are 9 months old, we use the mother’s response
to four questions, designed to capture her attitudes towards child rearing;
these include: importance for development of talking to the baby, cuddling
the baby, stimulating the baby, and importance of regular sleep and eating
time for the baby. For all the four questions mother responds on a …ve point
Likert scale from ‘strongly agree’to ‘strongly disagree’. As the children get
older we use variables that record the di¤erent ways that parents regulate
their child’s behaviour and their relationship with the child. These include:
whether the child has a regular bedtime; how much TV the child watches; and
whether the parents smack, or shout at the child if they are being naughty
(3 categories).
Melhuish et al. (2008) in their analysis …nd that the home learning environment is more important, relative to the social/routine activities, for the
cognitive development of the children. In our paper we do some robustness
16
check where we use only the home learning measures of parental inputs, and
drop the social/routine measures; the results (available on request) from this
re-de…ned latent parental investment are almost identical, qualitatively and
quantitatively, to the results presented in the paper.
4.3.2
Other characteristics
Variables in the vector Xt which a¤ect the latent cognitive ability are the
child’s age in months, and our main variable of interest, poverty status. It
is important to control carefully for age since the cognitive tests are typically standardised against norms within a three month age range, and hence
there may still be variation in cognitive development within these age groups
(Connelly, 2013).
In vector Xt we include a range of variables which, in the literature,
have been shown to a¤ect the parental inputs. It has been established
in the literature that larger families have a negative impact on children’s
educational outcomes, justi…cation for this comes from the resource (…nancial
and time) constraints hypothesis (Iacovou, 2001; Black et al., 2005). To
capture the resource constraints that the family might face we include two
variables: number of other siblings in the household; and a dummy for a
single parent household. We also include a poverty dummy in vector Xt ,
as parental investment is one of the pathways via which poverty impacts
children’s development (Duncan et al. 1998). Finally we include mother’s
education (= 1 if NVQ 4 or above, corresponding to higher education or
equivalent), to capture the fact that educated parents, especially mothers,
systematically spend (invest) more time in their children (Guryan et al.,
2008).
In vector X0 , which captures the initial conditions (both 0 and 0 ) we
use: birth weight, a dummy for the …rst born children (=1 if the child is …rst
born), mother’s age at birth, mother’s education at birth; and ethnicity of
the child (=1 for white children). We include birth weight as it is often used
as a proxy both for genetic endowment and for prenatal resource allocations
by parents (Del Bono et al., 2012). Birth order has been shown to be
signi…cant in long term outcomes for children (Black et al., 2005), with …rst
born children outperforming their younger siblings; to capture this we include
a dummy for …rst born children. Mother’s age and education at birth are
included to capture any early disadvantage that the child might face, as young
and less educated mothers often come from disadvantaged backgrounds which
they pass on to their children (Hawkes and Joshi, 2012). Using the MCS
data Dearden et al. (2006) have shown there are di¤erences in birth outcomes
(especially gestation and birth weight) by ethnicity and these di¤erences
17
remain even after controlling for various confounding factors; to capture these
di¤erence we include a dummy for ethnicity in our analysis.
5
Results
In this paper, the …nal sample for analysis comprises 8,741 children. These
are the children for whom we have information across all four sweeps of the
MCS; any loss of observations is due to attrition and missing information
on relevant covariates. For details of the sample used, see Appendix A.
Descriptive statistics (not weighted) for all the covariates and measurement
variables are given in Table A2 in Appendix A.
5.1
Episodic and persistent poverty
Table 1 reports the episodic incidence of child poverty in our sample according
to the measure described in sub-section 4.1. The incidence of poverty in the
sample is about 20% over the four sweeps. This is rather lower than the
incidence in the MCS sample when each sweep is examined separately, which
is as high as 30%. However the rate of child poverty we …nd is similar to
that reported in other studies using a balanced sample from the MCS (see,
for example, Schoon et al., 2012).
Table 2 presents the individual poverty pro…les, and the proportion of
children who experience each poverty pro…le. The interpretation is as follows.
For T = 2, there are 4 di¤erent poverty pro…les: P S = 00 indicates no
episodes of poverty while P S = 01 indicates that the child was not in poverty
in the …rst sweep but was in poverty in the second sweep etc.. Analogously,
for T = 4, there are 16 di¤erent poverty pro…les, and P S = 1111 denotes
being in poverty in all four sweeps. Finally, P P P is the prevalence of
persistent poverty. As can be seen, (100-64.1=) 36% of all children have
experienced at least one spell of relative poverty by the time they are aged
7. This is much higher than the 19% of children who are in poverty at age
7 (Table 1).
5.2
Child poverty and cognitive development
The correlations between the eight test scores are shown in Table 3. As can
be seen, children’s performance on each of the tests is positively correlated
with their ranking on other tests. Moreover, the tests would appear to capture di¤erent dimensions of cognitive development: for example, the highest
correlation in the ranking for BAS-PC(7) is the equivalent test taken two
18
years earlier, BAS-PC(5), rather than with any other test administered at
age 7.
Figure 2 shows the average test score ranking according to the poverty
status of the household at the time the test was taken. As can be clearly
seen, the average test scores for the non-poor children are signi…cantly higher
than the average scores for the children in poverty across all tests in all years.
This …nding is consistent with the previous literature in this area. Figure 3
shows the average scores for the two extreme poverty pro…les: children who
have never been in poverty and those that have always been in poverty at
each sweep. The di¤erences are larger here than in Figure 2 and this is
prima facie evidence to suggest that there may be cumulative e¤ects from
persistent poverty on cognitive test score outcomes.
There may be a number of possible explanations for the di¤erences observed in the raw data. For example: even though the test scores are
age-adjusted (within 3 month age bands), children’s cognitive development
is extremely rapid in their early years, and the tests are not administered to
all children at exactly the same age. Hence the age of the child when tested
can impact signi…cantly on the test score outcome. Also, as suggested by the
previous literature and the Field (2010) review, the background characteristics of the child, parental investment and parenting style may also in‡uence
the test scores. Our estimation method directly addresses these various
issues.
5.2.1
Baseline model
The results from the two di¤erent SEM speci…cations, of the baseline model,
are reported in Table 4 (for episodic poverty) and Table 5 (for persistent
poverty), respectively. What we have reported are the estimated coe¢ cients
of the structural model (equations (1) to (3)). The coe¢ cients (factor loadings) from the measurement models (equations (4) and (5)) are available on
request. Panel A in both tables reports estimates of equation (1) and Panel
B reports the estimated coe¢ cients of equation (2).
Findings consistent across the two speci…cations are as follows. From
Panel A of both Tables 4 and 5, there is evidence of clear persistence with
respect to cognitive ability ( t ) – previous latent cognitive ability is positively and signi…cantly correlated with current latent cognitive ability. Thus
a child developing well at age 3 is also likely to be doing well at age 5 and age
7, even after controlling for all other factors. From Table 5, Panel A, a one
standard deviation (SD) higher latent cognitive ability at age 3 is associated
with a 0.694 SD higher latent cognitive ability at age 5; this is equivalent
to 19 percentile ranks; similarly a one SD higher latent ability at age 5 is
19
associated with a 0.894 SD higher latent ability at age 7; equivalent to 25
percentile ranks. The percentile rank changes are calculated by multiplying
the observed SD changes in the latent variable by the SD of the underlying
measures; all the test scores have SD of around 28 (see Table A2 in Appendix). Also, higher birth weight, higher mothers age at the time of birth,
having a mother with higher education (NVQ 4 or above), being …rst born,
and being white are all associated with higher development at 9 months.
Panel A from both Tables 4 and 5, also reveals that lagged latent parental
investment ( t 1 ) has a positive and signi…cant impact on a child’s latent
cognitive ability, at all ages. From Table 5, if parental investment at age 3
increases by one SD, then the child’s cognitive ability at age 5 would increase
by 0.269 SD, equivalent to an increase of 8 percentile ranks. Finally, panel B
reveals that mothers with higher education on average provide higher levels
of parental investment at all ages while having other siblings in the household
signi…cantly reduces the level of parental investment in the child at age 3 and
age 5.
Current poverty has a negative and a signi…cant impact on cognitive development, at all ages, in both speci…cations. The only exception is the
insigni…cant e¤ect of current poverty on development at 9 months in speci…cation 2 reported in Table 5. Poverty at 9 months has a signi…cant e¤ect
on parental investment at 9 months (in both speci…cations) and at 3 years
(in speci…cation 2).
To capture fully the impact of persistent poverty on the cognitive ability
of the child we now focus on the estimates from speci…cation 2 as presented
in Table 5. At age 3, a child who is in poverty can be expected to be 0.263
SD (7 percentile ranks) below the latent cognitive ability score of a child who
has no experience of poverty. However, a child who has been persistently
in poverty since birth can be expected to be (0.263+0.189=) 0.452 SD (13
percentile ranks) below the latent cognitive ability score of the child who has
never been in poverty.
As noted above, one important bene…t of the SEM approach is that it
allows us to separately identify the direct and the indirect e¤ects of poverty
on latent cognitive ability. These are presented in Table 6, for the SEM
speci…cation 2 in Table 5. The interpretation of the table is as follows.
Reading down the …rst column, the total e¤ect of P 1 (poverty at 9 months,
or birth) on latent cognitive ability at age 3 is 0.254 SD. The direct e¤ect is
0.189 (the coe¢ cient on the poverty dummy in Table 5). However, there is
also an indirect e¤ect through the impact of P 1 on latent parental investment
in the child at 9 months, 1 , and through past cognitive development, 1 ,
which then a¤ects the child’s cognitive ability. The indirect via parental
investment is just –0.006 SD, and the indirect e¤ect via past development is
20
–0.059; both are not signi…cant. The total e¤ect of P 1 on latent cognitive
ability is then the sum of the direct and indirect e¤ects.
We can perform similar calculations at each age. For example, at age 7,
while the direct e¤ect of poverty at 9 months (P 1) on latent cognitive development is not statistically signi…cant, the indirect e¤ect of P 1 is signi…cantly
negative. These indirect e¤ects of poverty at 9 months on cognitive development at age 7 are manifested through latent parental investment, with e¤ects
from age 3 years being signi…cant, as well as through poorer past cognitive
ability, especially at age 3.
The direct e¤ect of being in persistent poverty on cognitive development
at age 7, if we just consider the statistically signi…cant e¤ects is: (–0.084–
0.098 =) –0.182 SD; 5 percentile ranks lower than the child who has never
been in poverty. However the total (direct + indirect) e¤ect of being persistently in poverty is (–0.232 –0.196 –0.149 –0.084 =) –0.661 SD; which translates to almost 19 percentile ranks lower than the child who has never been
in poverty. Three-quarters of the impact of being in poverty on children’s
cognitive development is driven by its indirect e¤ects on parental investment
and the persistence in cognitive ability (mainly the latter).
5.2.2
Model with endogenous inputs
The results for the model with endogenous inputs are presented in Tables 7,
8 and 9. In the …rst stage, we use the following proxies (Rt ) for t : number
of siblings in the households, single parent household, and whether or not
mother works. Justi…cation for these comes from the resource constraint
hypothesis mentioned in section 4.3.2 above. (Results for the …rst stage are
available on request.)
Table 7 gives the result for speci…cation 1 but this time taking into account the endogeneity of inputs; these results need to be compared with the
results in Table 4. In Panel B, in the model for latent parental investment
we now have current latent ability of the child (equation (10)). For all waves
the coe¢ cient on current cognitive ability of the child is insigni…cant. Overall the only di¤erence of taking into account endogeneity of inputs, relative
to the baseline model is that coe¢ cient on poverty dummy in latent parental
investment equation at age 9 months is now insigni…cant. Similarly comparing speci…cation 2 from the baseline model (Table 5) to the model with
endogeneity of inputs (Table 8) there is no change in results quantitatively
or qualitatively; the only exception being that poverty dummy in the equation for latent parental investment at 9 months (panel B, Table 8) is now
insigni…cant.
Our …nding is not at odds with the …nding in the literature. Cunha and
21
Heckman (2008) estimate a linear technology function as we do here and …nd
that allowing for endogeneity in inputs does not change their estimates from
the baseline model. Similarly Cunha et al. (2010) estimate a non-linear
technology function and …nd that allowing for endogeneity in inputs does
not change their estimates from the baseline model.
6
Concluding discussion
There is a consensus in the literature that family background, parental inputs, and income poverty can all have signi…cant e¤ects on children’s early
cognitive development. Much has been written about the importance of education and cognitive skills in early years for future life trajectories. Given
the high degree of persistence in ability formation, di¤erences in early year’s
ability are one of the main sources of variation in socioeconomic outcomes
across individuals.
This paper documents the impact of both episodic poverty and persistent
poverty on the cognitive development of children in the UK. Controlling for
parental investments and family circumstances, we …nd evidence of a direct
negative impact of income poverty on the cognitive development of children,
consistent with the recent evidence of Dahl and Lochner (2012) for the US.
Moreover, as in Schoon et al (2012), we …nd that persistent poverty has a
larger cumulative negative impact on children’s cognitive development than
episodic poverty.
As a robustness check, we experimented with using the log of household equivalised income instead of the poverty dummy in our model. The
estimated coe¢ cient on log income was positive indicating a non-linear relationship between ability formation and income. This …nding of diminishing
returns to income is also consistent with Dahl and Lochner (2012) who found
that the link between ability and income is stronger at lower levels of income.
Including log income instead of poverty status had no qualitative or quantitative impact on the other estimated coe¢ cients in our model (results available
on request).
Taking into account both the direct e¤ects and the indirect e¤ects of
poverty on parental investment, the cognitive development test scores for
children who are persistently in poverty throughout their early years are
almost 20 percentile ranks lower at age 7 than for children who have never
experienced poverty. This result is robust to the parental investment and
family background of the child. Given the evidence of strong persistence
in cognitive development, any detrimental impact of poverty on children’s
cognitive development in their early years is likely to have a lasting legacy
22
e¤ect well beyond the particular episodes of poverty. Poverty at birth and/or
age 3 can therefore seriously impact on children’s development by the time
they start school and thereafter well into their adult lives. This suggests
that policy targeted at poverty alleviation should be directed at these very
early years.
As in Schoon et al. (2012) and Kiernan and Menash (2011), we also …nd
that positive parenting can mitigate the impact of poverty to some extent.
Those who would argue that the quality of parenting skills and investment
are important for children’s development may therefore draw some encouragement from our results. It is clear that, controlling for income, parental
investments do indeed impact signi…cantly on children’s cognitive ability.
However, we …nd evidence that poverty also adversely a¤ects parental investments, especially in the very early years, and this subsequently has a
negative impact upon children’s cognitive development. Thus poverty not
only has a direct negative impact on children’s cognitive development, but
also has an indirect e¤ect through its adverse impact on parental inputs.
However, this result weakens when the endogeneity of inputs is taken into
account.
While we have focussed solely on cognitive skills, it has been established
in the literature that non-cognitive development is also important. For their
US data, Cunha et al. (2010) …nd that including non-cognitive ability lowers
the estimates of self-productivity i.e. the estimated coe¢ cient on lagged
cognitive ability. The impact of neglecting non-cognitive skills in our model
is therefore an empirical question for future work in this area.
23
Acknowledgements
We are grateful to The Centre for Longitudinal Studies, Institute of Education, for the use of, and to the UK the Data Archive and Economic and
Social Data Service for providing access to, the data from the Millennium
Cohort Study. We are also grateful to Lucinda Platt in particular for her assistance with the poverty data. We also thank Sarah Brown, Aki Tsuchiya,
Arne Risa Hole, Lucinda Platt and participants at the Work, Pensions and
Labour Economics Study Group (WPEG) conference, She¢ eld, July 2011;
at the Society for the Study of Economic Inequality (ECINEQ) conference,
Catania, Sicily, July 2011; at a Department for Work and Pensions (DWP)
seminar, London, September 2011; at the Canadian Economic Association
conference, Calgary, Canada, June 2012; and at the Delhi School of Economics, Delhi, India, August 2014,for their helpful comments. We would
also like to thank two anonymous refrees and the joint editor for their very
valuable comments, which have much improved the paper. All responsibility
for the analysis and interpretation of these data lies with the authors.
24
References
1. Barnes, M., Chanfreau, J. and Tomaszewski, W. (2010) Growing up in
Scotland: The Circumstances of Persistently Poor Children.
http://www.scotland.gov.uk/Publications/2010/04/26095519/0
2. Becker, G. and Tomes, N. (1986) Human Capital and the Rise and Fall
of Families. Journal of Labor Economics, 4(3), S1-S39.
3. Black, S.E., Devereux, P.J. and Salvanes, K.G. (2005). The more the
merrier? The e¤ect of family size and birth order on children’s education. The Quarterly Journal of Economics, 120(2), 669-700.
4. Blanden, J., Gregg, P. and Macmillan, L. (2007) Accounting for intergenerational income persistence: non-cognitive skills, ability and education. Economic Journal, 117, C43-C60.
5. Blanden, J. and Machin, S. (2010) Changes in inequality and intergenerational mobility in early years assessments. In Children of the 21st
century: The …rst …ve years (eds. K. Hansen, H. Joshi and S. Dex),
chapter 9. Bristol: Policy Press.
6. Bossert, W., Chakravarty, S.R. and d’Ambrosio, C. (2012). Poverty
and time. The Journal of Economic Inequality, 10(2), 145-162.
7. Bracken, B.A. (2002) Bracken School Readiness Assessment: Administration Manual. San Antonio, Texas: Psychological Corporation.
8. Brooks-Gunn, J., Duncan, G.J., Klebanov, P.K., and Sealand, N. (1993)
Do neighborhoods in‡uence child and adolescent behavior? American
Journal of Sociology, 99(2), 353–95.
9. Brooks-Gunn, J. and Duncan, G.J. (1997) The e¤ects of poverty on
children. The Future of Children, 7(2), 55-71.
10. Carneiro, P. and Heckman, J.J. (2002) The evidence on credit constraints in post-secondary schooling. Economic Journal, 112(482), 705734.
11. Cameron, S.V and Heckman, J.J. (2001) The dynamics of educational
attainment for black, Hispanic, and white males. Journal of Political
Economy, 109(3), 455-99.
25
12. Connelly, R. (2013) Interpreting Test Scores. Millennium Cohort Study
Data Note 2013/1, Centre for Longitudinal Studies, Institute of Education, London.
13. Corcoran, M. (1995). Rags to Rags: Poverty and Mobility in the United
States. Annual Review of Sociology, 21, 237-267.
14. Cunha, F. and Heckman, J.J. (2008) Formulating, Identifying and Estimating the Technology of Cognitive and Noncognitive Skill Formation.
Journal of Human Resources, XLIII, 738-782.
15. Cunha, F., Heckman, J.J. and Schennach, S.M. (2010). Estimating the
technology of cognitive and noncognitive skill formation. Econometrica, 78(3), 883-931.
16. Cunha, F. (2011). Recent Developments in the Identi…cation and Estimation of Production Functions of Skills. Fiscal Studies, 32(2), 297316.
17. Dahl, G. and Lochner, L. (2012) The impact of family income on child
achievement: evidence from the earned income tax credit. American
Economic Review, 102(5), 1927-56.
18. Dearden, L., Mesnard, A. and Shaw, J. (2006). Ethnic Di¤erences in
Birth Outcomes in England. Fiscal Studies, 27(1), 17-46.
19. Del Bono, E., Ermisch, J. and Francesconi, M. (2012). Intrafamily resource allocations: a dynamic structural model of birth weight. Journal
of Labor Economics, 30(3), 657-706.
20. Dezateux, C., Bedford, H., Cole, T., Peckham, C., Schoon, I., Hope, S.
and Butler, N. (2004) Babies’health and development. In Millennium
Cohort Study First Survey: A User’s Guide to Initial Findings (eds. S.
Dex and H. Joshi), chapter 7. Centre for Longitudinal Studies, Institute
of Education, London.
21. Duncan, G.J., Yeung, W., Brooks-Gunn, J. and Smith, J.R. (1998)
How much does childhood poverty a¤ect the life chances of children?
American Sociological Review, 63, 406-423.
22. Duncan, G.J., Kathleen M.Z. and Kalil, A. (2010) Early-Childhood
Poverty and Adult Attainment, Behavior, and Health. Child Development, 81(1), 306-325.
26
23. Dutta, I., Roope, L. and Zank, H. (2013). On intertemporal poverty
measures: the role of a- uence and want. Social Choice and Welfare,
41(4), 741-762.
24. Elliott, C.D., Smith, P. and McCulloch, K. (1996) British Ability Scales
Second Edition (BAS II): Administration and Scoring Manual. London: NFER-Nelson.
25. Elliott, C.D., Smith, P. and McCulloch, K. (1997) British Ability Scales
Second Edition (BAS II): Technical Manual. London: NFER-Nelson.
26. Ermisch, J. (2008) Origins of social immobility and inequality: parenting and early childhood development. National Institute Economic
Review, 205(62), 62-71.
27. Feinstein, L. (2003) Inequality in the Early Cognitive Development of
British Children in the 1970 Cohort. Economica, 70, 73-97.
28. Field, F. (2010). The Foundation Years: preventing poor children becoming poor adults, The report of the Independent Review on Poverty
and Life Chances. The Stationery O¢ ce.
29. Foster, J. (2009) A Class of Chronic Poverty Measures. In Poverty
dynamics: Interdisciplinary perspectives (eds. A. Addison, D. Hulme
and R. Kanbur), chapter 3. Oxford: Oxford University Press.
30. Frankenburg, W.K. and Dodds, J. (1967) The Denver Developmental
Screening Test. The Journal of Pediatrics, 71 (2), 181-191.
31. Frankenburg, W.K., Dodds, J., Archer, P., Shapiro, H. and Bresnick,
B. (1992). The Denver II: a major revision and restandardization of
the Denver Developmental Screening Test. Pediatrics, 89(1), 91-97.
32. Gelber, A. and Isen, A. (2013) Children’s schooling and parents’ behavior: Evidence from the Head Start Impact Study. Journal of Public
Economics, 101, 25-38.
33. Goodman, A. and Gregg, P. (2010) Poorer children’s educational attainment: how important are attitudes and behaviour? Report for the
Joseph Rowntree Foundation.
http://www.jrf.org.uk/publications/educational-attainment-poor-children
34. Gregg, P. and Macmillan, L. (2009) Family income and education in the
next generation: exploring income gradients in education for current
cohorts of youth. CMPO Working Paper 09/223, University of Bristol.
27
35. Guryan, J., Hurst, E. and Kearney, M. (2008). Parental Education
and Parental Time with Children. Journal of Economic Perspectives,
22(3), 23-46.
36. Hansen, K. (ed.) (2012) Millennium Cohort Study: First, Second,
Third and Fourth Surveys –A Guide to the Datasets, Seventh Edition.
Centre for Longitudinal Studies, Institute of Education, London.
37. Hawkes, D. and Joshi, H. (2012). Age at motherhood and child development: Evidence from the UK Millennium Cohort. National Institute
Economic Review, 222(1), R52-R66.
38. Hawkes, D and Lewis, I. (2008) Missing Income Data in the Millennium
Cohort Study: Evidence from the First Two Sweeps. CLS Working
Paper, London: Centre for Longitudinal Studies.
39. HBAI (2010) Households below average income: an analysis of the
income distribution 1994/95-2008/09, Department for Work and Pensions, 20 May 2010.
40. Heckman, J.J and Masterov, D.V. (2007) The productivity argument
for investing in young children. Review of Agricultural Economics, 29(3),
446-493.
41. Iacovou, M. (2001) Family Composition and Children’s Educational
Outcomes. ISER Working Paper.
42. Jerrim, J. and Vignoles, A. (2011) Social mobility, regression to the
mean and the cognitive development of high ability children from disadvantaged homes. Journal of the Royal Statistical Society: Series A,
176(4), 887-906.
43. Ketende, S. (2010) Millennium Cohort Study Technical Report on Response, Third Edition. London: Centre for Longitudinal Studies, Institute of Education, University of London.
44. Kiernan, K.E and Huerta, M.C. (2008) Economic deprivation, maternal
depression, parenting and children’s cognitive and emotional development in early childhood. British Journal of Sociology, 59(4), 783-806.
45. Kiernan, K.E and Mensah, F.K. (2009) Poverty, maternal depression,
family status and children’s cognitive and behavioural development in
early childhood: A longitudinal study. Journal of Social Policy, 38,
569-588.
28
46. Kiernan, K.E and Mensah, F.K. (2011) Poverty, family resources and
children’s early educational attainment: the mediating role of parenting. British Educational Research Journal, 37(2), 317-336.
47. Knudsen, E.I., Heckman, J.J., Cameron, J.L. and Shonko¤, J.P. (2006).
Economic, neurobiological, and behavioral perspectives on building
America’s future workforce. Proceedings of the National Academy of
Sciences, 103(27), 10155-10162.
48. Leibowitz, A. (1974) Home Investments in Children. Journal of Political Economy, 82(2), S111-31.
49. Mayer, S. (1997) What money can’t buy: Family income and children’s
life chances. Harvard University Press. Cambridge, MA.
50. Melhuish, E.C., Phan, M.B., Sylva, K., Sammons, P., Siraj-Blatchford,
I. and Taggart, B. (2008). E¤ects of the home learning environment and
preschool center experience upon literacy and numeracy development
in early primary school. Journal of Social Issues, 64(1), 95-114.
51. Muthén, L.K. and Muthén, B.O. (2010). Mplus User’s Guide, Sixth
Edition. Los Angeles, CA: Muthén & Muthén.
52. OECD (2009) What are equivalence scales?
http://www.oecd.org/eco/growth/OECD-Note-EquivalenceScales.pdf
53. Pakpahan, Y.M., Suryadarma, D. and Suryahadi, A. (2009). Destined for destitution: intergenerational poverty persistence in Indonesia. Chronic poverty research centre (CPRC).
54. Plug, E. and Vijverberg, W. (2005) Does Family Income Matter for
Schooling Outcomes? Using Adoptees as a Natural Experiment. The
Economic Journal, 115, 879-906.
55. Schoon, I., Cheng, H. and Jones, E (2010) Resilience in children’s development. In Children of the 21st century: The …rst …ve years (eds.
K. Hansen, H. Joshi and S. Dex), chapter 14. Bristol: Policy Press.
56. Schoon, I., Jones, E., Cheng, H. and Maughan, B. (2012). Family hardship, family instability and children’s cognitive development. Journal
of Epidemiology and Community Health, 66 (8), 716-722
57. Stewart, M.B. (1983) On least-squares regression when the dependent
variable is grouped. Review of Economic Studies, 50, 737-753.
29
58. Todd, P.E. and Wolpin, K.I. (2003) On the speci…cation and estimation of the production function for cognitive achievement. Economic
Journal, 113 (485), 3-33.
59. Todd, P.E. and Wolpin, K.I. (2007) The production of cognitive achievement in children: Home, school, and racial test score gaps. Journal of
Human Capital, 1(1), 91-136.
60. Waldfogel, J. (2010) Britain’s War on Poverty, New York: Russell Sage
Foundation.
30
Figure 1: Structural and measurement model
Yθt-1,1
Measurement
Model
Yθt-1,2
Yθt,1
Measurement
Model
Yθt,2
θt
θt-1
Xλ
λt-1
Xθ
Structural Model
Yλt-1,1
Yλt-1,2
Measurement Model
Notes: The figure illustrates the dynamics of the structural and the measurement models for the baseline specification (i.e.
the specification without endogenous inputs), outlined in section 3 of the paper, with two time periods; in the paper we have
four time periods and initial conditions. The unobservable (latent) variables are in ellipses and the observable variables in
rectangles. The solid rectangle gives us the structural relationship (given by equations (1) and (2) in the paper) and the
dotted rectangles represent the measurement models (given by equations (4) and (5) in the paper). The single headed
arrows illustrate the theorised unidirectional causal relationships between variables. For the structural model: ability at time
t (θt) is determined by past ability (θt-1), past parental investment (λt-1), and covariates (Xθ); parental investment is in turn
influenced by covariates Xλ. If we take the first two time periods (i.e. t = 2) then there will be initial conditions represented
by X0, which will influence θ1. Each unobservable latent variable (θ and λ) is measured by a series of observable variables
θ
λ
(Y and Y , respectively); this measurement model will vary over time.
31
Figure 2: Average test rank scores by poverty state: period-by-period
70
Average test score
60
60
60
58
50
40
40
46
42
57
56
53
46
45
55
55
45
42
44
30
20
10
0
BSRA (3)
BAS-NV BAS-PS (5) BAS-NV BAS-PC (5) BAS-WR BAS-PC (7) PiM (7)
(3)
(5)
(7)
not in poverty
in poverty
Figure 3: Average test rank scores by poverty state: never vs always in poverty
70
62
60
Average test score
62
60
54
50
40
59
57
44
38
42
38
57
57
38
41
40
33
30
20
10
0
BSRA (3) BAS-NV (3) BAS-PS (5) BAS-NV (5) BAS-PC (5) BAS-WR BAS-PC (7) PiM (7)
(7)
never in poverty
always in poverty
32
Table 1: Poverty incidence
Sweep
Average age of the child
Poverty rate
Sample size
MCS1
2001-2
9 months
MCS2
2004-5
3 years
MCS3
2006
5 years
MCS4
2008
7 years
20.2
8,741
21.2
8,741
21.5
8,741
18.7
8,741
Notes to Table 1:
1. Poverty rate is based on the poverty indicators provided by the MCS (see text for further inforamtion). The
threshold is household equivalised income less than 60% of median household income where income is
equivalised according to the OECD equivalence scale.
2. In all reported statistics, the MCS weights which take into account the survey design, non-response bias and
attrition over time have been taken into account.
Table 2: The prevalence of persistent poverty
Row
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
Time horizon T = 2
PS
PPP
00
72.1
01
7.7
10
6.7
11
13.5
Time horizon T = 3
PS
PPP
000
67.0
001
5.1
010
4.1
100
4.4
101
2.3
011
3.6
110
2.9
111
10.5
100.0%
100.0%
Time horizon T = 4
PS
PPP
0000
64.1
0001
2.9
0010
3.5
0100
3.4
1000
3.8
0101
0.7
1001
0.6
1010
1.1
0011
1.6
0110
1.4
1100
1.8
1011
1.2
1101
1.2
0111
2.2
1110
2.4
1111
8.2
100.0%
Notes to Table 2:
1. PS is the poverty profile or status. The digits describe the poverty status in each sweep, so that, for example,
001 represents individuals who were not in poverty in MCS1 nor in MCS2 but are in poverty in MCS3 – see
text for details.
2. PPP is prevalence of persistent poverty (i.e. the proportion of the sample in each poverty state) calculated
using the poverty rate measure reported in Table 1.
3. In all reported statistics, the MCS weights which take into account the survey design, non-response bias and
attrition over time have been taken into account.
33
Table 3: Cognitive assessment scores
MCS2: age 3
1
2
Test
1
2
3
4
5
6
7
8
BSRA(3)
BSRA(3)
3
MCS3: age 5
4
5
6
MCS4: age 7
7
BAS-NV(3) BAS-PS(5) BAS-NV(5) BAS-PC(5) BAS-WR(7) BAS-PC(7)
BAS-NV(3)
1.000
0.557
1.000
BAS-PS(5)
0.270
0.206
1.000
BAS-NV(5)
0.489
0.493
0.295
1.000
BAS-PC(5)
0.328
0.245
0.319
0.317
1.000
BAS-WR(7)
BAS-PC(7)
0.435
0.325
0.298
0.247
0.230
0.300
0.359
0.290
0.324
0.552
1.000
0.317
1.000
PiM(7)
0.370
0.262
0.307
0.365
0.383
0.490
0.468
8
PiM(7)
1.000
Notes to Table 3:
1. The tests scores are:
BSRA(3) – percentile rank Bracken School Readiness Assessment, age 3
BAS-NV(3) – percentile rank BAS naming vocabulary, age 3
BAS-PS(5) – percentile rank BAS picture similarity, age 5
BAS-NV(5) – percentile rank BAS naming vocabulary, age 5
BAS-PC(5) – percentile rank BAS pattern construction, age 5
BAS-WR(7) – percentile rank BAS word reading, age 7
BAS-PC(7) – percentile rank BAS pattern construction, age 7
PiM(7) – percentile rank Progress in Maths, age 7
2. Correlations are (weighted) pairwise Pearsonian correlations between the percentile ranking on each test.
3. All correlations are statistically significantly different from zero at the p = 0.01 level.
4. In all reported statistics, the MCS weights which take into account the survey design, non-response bias and
attrition over time have been taken into account.
34
Table 4: Specification 1 – Latent cognitive development and the incidence of poverty
Panel A
Latent cognitive development 𝜃𝑡
MCS2: age 3 years MCS3: age 5 years
𝜃𝑡−1
𝜆𝑡−1
𝑃𝑡
Age (months)
Initial conditions
Birth weight
First born
Mother’s age (years)
Mother’s education
Ethnicity: white
MCS1: age 9
MCS4: age 7
months
years
𝜃1
𝜃2
𝜃3
𝜃4
Coefficient p-value Coefficient p-value Coefficient p-value Coefficient p-value
0.000
0.000
0.000
0.504
0.700
0.908
0.001
0.000
0.000
0.072
0.267
0.076
-0.468
0.000
0.000 -0.132
-0.249
0.005 -0.087
0.070
0.017
0.450
0.024
0.123 -0.055
0.032
0.076
0.003
0.287
0.195
0.197
0.391
0.463
0.000
0.009
0.000
0.001
0.000
-
-
-
-
-
-
Panel B
Latent parental investment 𝜆𝑡
MCS1: age 9 months
MCS2: age 3 years
MCS3: age 5 years
𝜆1
𝜆2
𝜆3
effect
p-value
effect p-value
effect p-value
-0.095
0.071
-0.016
0.722
-0.025
0.597
𝑃𝑡
Mother’s education
0.148
0.107
0.244
0.006
0.160
0.025
Other siblings
0.008
0.853
-0.108
0.013
-0.169
0.000
Single parent household 0.041
0.549
0.012
0.875
-0.102
0.089
Notes to Table 4:
1. Panel A gives the estimates of the structural parameters in equation (1) in the paper, where latent ability at
time t (θt) is determined by past latent ability (θt-1), past latent parental investment (λt-1), and a set of control
variables including poverty at time t (Pt). For the first period, t=1, we have a set of covariates which capture
the initial conditions (X0); equation (3) in the paper. Panel B gives the estimates of the structural parameters
in equation (2) in the paper, where parental investment at time t (λt) is determined by a set of control
variables including poverty at time t (Pt).
2. All the reported coefficients are standardized. For the continuous independent variables, the coefficient
represents the change in the dependent variable associated with a one standard deviation (SD) change in the
independent variable. For the binary independent variables the coefficient represents the change associated
with a shift in the variable from 0 to 1.
3. In all reported statistics, the MCS weights which take into account the survey design, non-response bias and
attrition over time have been taken into account.
4. Sample size: 8,741; CFI = 0.714; RMSE = 0.029.
35
Table 5: Specification 2 – Latent cognitive development and the persistence of poverty
Panel A
𝜃𝑡−1
𝜆𝑡−1
𝑃𝑡
𝑃𝑡−1
𝑃𝑡−2
𝑃𝑡−3
Age
Initial conditions
Birth weight
First born
Mother’s age
Mother’s education
Ethnicity: white
MCS1: age 9 months
𝜃1
Coefficient p-value
-0.131
0.174
0.015
0.515
0.311
0.208
0.192
0.402
0.457
Latent cognitive development 𝜃𝑡
MCS2: age 3 years MCS3: age 5 years MCS4: age 7 years
𝜃2
𝜃3
𝜃4
Coefficient p-value Coefficient p-value Coefficient p-value
0.000
0.000
0.000
0.449
0.694
0.894
0.001
0.000
0.000
0.072
0.269
0.076
0.000 -0.099
-0.263
0.056 -0.084
0.077
0.004
-0.189
0.076
0.148 -0.058
0.209
-0.058
0.216 -0.098
0.030
-0.001
0.985
0.024
0.115 -0.055
0.032
0.076
0.004
0.000
0.008
0.000
0.001
0.000
-
-
-
-
-
-
Panel B
Latent parental investment 𝜆𝑡
MCS1: age 9 months
MCS2: age 3 years
MCS3: age 5 years
𝜆1
𝜆2
𝜆3
effect
p-value
effect p-value
effect p-value
-0.087
0.099
-0.021
0.659
-0.024
0.619
𝑃𝑡
-0.084
0.082
0.026
0.563
𝑃𝑡−1
-0.013
0.776
𝑃𝑡−2
Mother’s education
0.150
0.101
0.250
0.005
0.161
0.025
Other siblings
0.008
0.852
-0.108
0.013
-0.169
0.000
Single parent household 0.041
0.549
0.012
0.872
-0.102
0.091
Notes to Table 5:
1. Refer to notes to Table 4.
2. Sample size: 8,741; CFI = 0.713; RMSE = 0.029.
36
Table 6: Identifying direct and indirect effects of persistent poverty on cognitive development
Effects from P1:
Total effect on 𝜃𝑡
Direct effect on 𝜃𝑡
Indirect effect on 𝜃𝑡
Indirect via 𝜆1
Indirect via 𝜆2
Indirect via 𝜆3
Indirect via 𝜃1
Indirect via 𝜃2
Indirect via 𝜃3
Effects from P2:
Total effect on 𝜃𝑡
Direct effect on 𝜃𝑡
Indirect effect on 𝜃𝑡
Indirect via 𝜆2
Indirect via 𝜆3
Indirect via 𝜃2
Indirect via 𝜃3
Effects from P3:
Total effect on 𝜃𝑡
Direct effect on 𝜃𝑡
Indirect effect on 𝜃𝑡
Indirect via 𝜆3
Indirect via 𝜃3
Effects from P4:
Total effect on 𝜃𝑡
Direct effect on 𝜃𝑡
Latent cognitive development 𝜃𝑡
MCS2: age 3
MCS3: age 5
MCS4: age 7
𝜃2
𝜃3
𝜃4
effect p-value
effect p-value
effect p-value
-0.254
-0.189
-0.065
-0.006
0.000
0.004
0.131
0.148
-0.257
-0.058
-0.199
-0.004
-0.023
0.000
0.216
0.000
0.145
0.088
-0.059
0.173
-0.041
-0.131
0.177
0.004
-0.263
-0.263
0.000
0.000
-0.112
0.076
-0.188
-0.006
0.028
0.148
0.000
0.659
-0.182
0.000
-0.099
-0.099
0.056
0.056
Notes to Table 6:
1. Refer to notes to Table 4.
2. P1 – poverty status at 9 months
P2 – poverty status at 3 years
P3 – poverty status at 5 years
P4 – poverty status at 7 years
37
-0.232
-0.001
-0.231
-0.004
-0.020
-0.001
-0.037
-0.117
-0.052
0.000
0.985
0.000
0.147
0.088
0.776
0.176
0.004
0.217
-0.196
-0.098
-0.098
-0.005
0.002
-0.163
0.068
0.000
0.030
0.032
0.659
0.569
0.000
0.147
-0.149
-0.058
-0.090
-0.002
-0.088
0.003
0.209
0.056
0.623
0.057
-0.084
-0.084
0.077
0.077
Table 7: Specification 1 – Latent cognitive development and the incidence of poverty
(Endogenous inputs)
Panel A
𝜃𝑡−1
𝜆𝑡−1
𝑃𝑡
Age
Initial conditions
Birth weight
First born
Mother’s age
Mother’s education
Ethnicity: white
MCS1: age 9 months
𝜃1
Coefficient p-value
-0.462
0.000
0.018
0.412
0.264
0.184
0.201
0.373
0.471
0.000
0.009
0.000
0.001
0.000
Latent cognitive development 𝜃𝑡
MCS2: age 3 years MCS3: age 5 years MCS4: age 7 years
𝜃2
𝜃3
𝜃4
Coefficient p-value Coefficient p-value Coefficient p-value
0.000
0.000
0.000
0.557
0.693
0.905
0.001
0.000
0.000
0.072
0.266
0.076
0.000 -0.131
-0.246
0.006 -0.086
0.072
0.025
0.106 -0.055
0.033
0.076
0.004
-
-
-
-
-
-
Panel B
𝑃𝑡
𝜃𝑡
Mother’s education
Other siblings
Single parent household
Latent parental investment 𝜆𝑡
MCS1: age 9 months
MCS2: age 3 years
𝜆1
𝜆2
effect
p-value
effect p-value
-0.079
0.135
-0.029
0.527
-0.161
0.138
0.142
0.123
0.254
0.004
0.025
0.656
-0.108
0.018
0.054
0.638
-0.095
0.394
MCS3: age 5 years
𝜆3
effect p-value
-0.033
0.494
-0.032
0.673
0.161
0.025
-0.171
0.000
-0.098
0.102
Notes to Table 7:
1. Panel A gives the estimates of the structural parameters in equation (1) in the paper, where latent ability at
time t (θt) is determined by past latent ability (θt-1), past latent parental investment (λt-1), and a set of control
variables including poverty at time t (Pt). For the first period, t=1, we have a set of covariates which capture
the initial conditions (X0); equation (3) in the paper. Panel B gives the estimates of the structural parameters
in equation (10) in the paper, where parental investment at time t (λt) is determined by ability at time t (θt)
and a set of control variables including poverty at time t (Pt).
2. All the reported coefficients are standardized. For the continuous independent variables, the coefficient
represents the change in the dependent variable associated with a one standard deviation (SD) change in the
independent variable. For the binary independent variables the coefficient represents the change associated
with a shift in the variable from 0 to 1.
3. Sample size: 8,741; CFI = 0.717; RMSE = 0.028.
4. In all reported statistics, the MCS weights which take into account the survey design, non-response bias and
attrition over time have been taken into account.
38
Table 8: Specification 2 – Latent cognitive development and the persistence of poverty
(Endogenous inputs)
Panel A
𝜃𝑡−1
𝜆𝑡−1
𝑃𝑡
𝑃𝑡−1
𝑃𝑡−2
𝑃𝑡−3
Age
Initial conditions
Birth weight
First born
Mother’s age
Mother’s education
Ethnicity: white
MCS1: age 9 months
𝜃1
Coefficient p-value
-0.094
0.361
0.016
0.500
0.312
0.206
0.192
0.400
0.459
Latent cognitive development 𝜃𝑡
MCS2: age 3 years MCS3: age 5 years MCS4: age 7 years
𝜃2
𝜃3
𝜃4
Coefficient p-value Coefficient p-value Coefficient p-value
0.000
0.000
0.000
0.449
0.698
0.894
0.001
0.000
0.000
0.072
0.268
0.075
-0.263
0.000 -0.097
0.064 -0.086
0.072
-0.215
0.002
0.083
0.123 -0.062
0.184
-0.047
0.342 -0.101
0.026
-0.011
0.809
0.025
0.113 -0.055
0.033
0.076
0.004
0.000
0.008
0.000
0.001
0.000
-
-
-
-
-
-
Panel B
𝑃𝑡
𝑃𝑡−1
𝑃𝑡−2
𝜃𝑡
Mother’s education
Other siblings
Single parent household
Latent parental investment 𝜆𝑡
MCS1: age 9 months
MCS2: age 3 years
𝜆1
𝜆2
effect
p-value
effect p-value
-0.073
0.170
-0.035
0.475
-0.116
0.019
-0.161
0.138
0.146
0.112
0.264
0.003
0.026
0.651
-0.108
0.018
0.054
0.638
-0.095
0.391
Notes to Table 8:
1. Refer to notes to Table 7.
2. Sample size: 8,741; CFI = 0.716; RMSE = 0.028.
39
MCS3: age 5 years
𝜆3
effect p-value
-0.031
0.517
0.021
0.636
-0.027
0.563
-0.032
0.673
0.162
0.025
-0.171
0.000
-0.098
0.102
Table 9: Identifying direct and indirect effects of persistent poverty on cognitive development
(Endogenous inputs)
Effects from P1:
Total effect on 𝜃𝑡
Direct effect on 𝜃𝑡
Indirect effect on 𝜃𝑡
Indirect via 𝜆1
Indirect via 𝜆2
Indirect via 𝜆3
Indirect via 𝜃1
Indirect via 𝜃2
Indirect via 𝜃3
Effects from P2:
Total effect on 𝜃𝑡
Direct effect on 𝜃𝑡
Indirect effect on 𝜃𝑡
Indirect via 𝜆2
Indirect via 𝜆3
Indirect via 𝜃2
Indirect via 𝜃3
Effects from P3:
Total effect on 𝜃𝑡
Direct effect on 𝜃𝑡
Indirect effect on 𝜃𝑡
Indirect via 𝜆3
Indirect via 𝜃3
Effects from P4:
Total effect on 𝜃𝑡
Direct effect on 𝜃𝑡
Latent cognitive development 𝜃𝑡
MCS2: age 3
MCS3: age 5
MCS4: age 7
𝜃2
𝜃3
𝜃4
effect p-value
effect p-value
effect p-value
-0.262
-0.215
-0.047
-0.005
0.000
0.002
0.032
0.210
-0.261
-0.047
-0.214
-0.004
-0.031
0.000
0.342
0.000
0.212
0.022
-0.042
0.036
-0.029
-0.150
0.362
0.002
-0.263
-0.263
0.000
0.000
-0.109
0.083
-0.192
-0.009
0.035
0.123
0.000
0.477
-0.183
0.000
-0.097
-0.097
0.064
0.064
Notes to Table 9:
1. Refer to notes to Table 7.
2. P1 – poverty status at 9 months
P2 – poverty status at 3 years
P3 – poverty status at 5 years
P4 – poverty status at 7 years
40
-0.246
-0.011
-0.235
-0.003
-0.028
-0.002
-0.026
-0.134
-0.042
0.000
0.809
0.000
0.211
0.022
0.566
0.360
0.002
0.342
-0.197
-0.101
-0.096
-0.008
0.002
-0.164
0.074
0.000
0.026
0.039
0.476
0.640
0.000
0.125
-0.152
-0.062
-0.089
-0.002
-0.087
0.003
0.184
0.064
0.524
0.067
-0.086
-0.086
0.072
0.072
Appendix A: Sample construction
Table A1: Productive Interviews
MCS Sweep
1
2
3
4
Number of
Households
18,552
15,590
15,246
13,857
Number of
Children
18,818
15,808
15,459
14,043
MCS surveyed 18,552 households (with 18,818 children, including twins and triplets) for the first
sweep, which were then to be followed over time. In the second sweep they added 692 additional
households (referred to as the ‘new families’) which should have been eligible for the first sweep
but were missed. The productive households (i.e. the households that completed the survey) at
each sweep are shown in the table above. Of the 18,552 families which were productive at sweep
1, only 11,721 (61%) were productive in all four sweeps. Refusing to participate was the biggest
reason for attrition (other reasons being: emigration, failure to contact, death, etc.); though there
are families which have declined to be interviewed in one sweep but then participated in
subsequent sweeps. The refusal rates are higher for the ‘disadvantaged’ and ‘ethnic minority’
families, relative to the ‘advantaged’ families, across all the four countries of the UK. For further
details on the response rates in the MCS, see Ketende (2010). Once a family has participated in
the survey, non-response on specific questions is low, with the exception of questions relating to
‘family income’; and this non-response is not random. Analysis by Hawkes and Lewis (2008)
shows that self-employed and families with low income are more likely to ‘refuse to’ answer the
questions related with income in MCS.
For our analysis we need information on children from all four sweeps, so our starting sample is
the 11,721 families which were productive in all four sweeps of MCS. By definition this will
exclude the ‘new families’ which were added in the second sweep. Further in our analysis we
use only one child per family (of the twins and triplets, we keep only one child). Any further loss
of observations is because of non-response on the relevant covariates used in our analysis. The
final sample we work with is 8,741 children, from 8,741 families.
41
Table A2: Unweighted descriptive statistics for the variables used in the analysis
MCS1: 9 months
Mean
SD
Measures for cognitive ability:
Gross motor function delay 1 (1 = delay)
Fine motor function delay 1 (1 = delay)
Communicative gestures delay 1 (1 = delay)
BSR composite score
BAS Naming Vocabulary
BAS Picture Similarity
BAS Pattern Construction
BAS Word Reading
Progress in Maths
Measures for Parental Investment and Style:
Importance of stimulating the baby 2
Importance of talking to baby 2
Importance of cuddling baby 2
Importance of regular sleeping/eating for baby 2
Mother reads to the child 3
Father reads to the child 3
Child paints/draws at home 3
Child helped with alphabet/reading 3
Child helped with writing 3
Child taken to the library 4
Child helped with counting/maths 3
Regular bedtime 5
Watching TV 6
Shout at the child 7
Smack the child 7
Other covariates:
Age of the child in months
0.121
0.134
0.399
MCS2: 3 years
Mean
SD
9.18
MCS4: 7 years
Mean
SD
53.5
53.3
1.373
1.166
1.167
1.573
MCS3: 5 years
Mean
SD
28.3
28.1
53.5
51.3
52.2
28.0
28.4
28.6
51.8
51.6
51.7
28.6
28.5
28.3
0.586
0.401
0.423
0.714
0.50
42
3.336
2.060
3.168
2.147
0.987
1.397
0.926
1.383
0.495
3.184
2.159
1.924
1.628
0.821
0.747
1.029
0.893
0.649
0.541
0.668
37.49
2.26
3.276
1.990
1.980
3.388
2.692
0.446
2.819
2.502
1.913
1.712
0.681
62.62
0.915
1.328
0.980
0.906
1.190
0.651
1.136
0.775
0.637
0.506
0.671
2.92
86.81
2.99
Mother’s education1 (1= NVQ 4 and higher)
Number of other siblings in the house
Single parent household1 (1 = yes)
Mother works1 (1 = yes)
Initial Conditions:
Ethnicity1 (1 = white; 0 = otherwise)
Birth weight (kgs)
Mother’s age at birth of the cohort child (years)
First born1 (1 = if the cohort child is first born)
MCS1: 9 months
Mean
SD
0.367
0.862
0.979
0.128
0.555
0.894
3.382
29.05
0.431
MCS2: 3 years
Mean
SD
0.367
1.128
0.996
0.149
0.581
-
MCS3: 5 years
Mean
SD
0.403
1.315
0.996
0.169
0.630
-
0.581
5.704
-
Notes to Table A2:
1. dummy variable
2. coded as: 1 = strongly agree, 2 = agree, 3 = neither agree nor disagree, 4 = disagree, 5 = strongly disagree
3. coded as: 0 = not at all, 1 = once/twice/less a month, 2 = once/twice a week, 3 = several times a week, 4 = every day
4. coded as: 0 = never/special occasions, 1 = at least once a month, 2 = once a week or more
5. coded as: 0 = never or almost never, 1 = sometimes, 2 = usually, 3 = always
6. coded as: 0 = none, 1 = up to one hour a day, 2 = between 1 and 3 hours a day, 3 = more than three hours a day
7. coded as: 0 = never, 1 = rarely, 2 = more than once a month
43
MCS4: 7 years
Mean
SD